neuron mcp — MCP Server for Hook-less Editors

neuron mcp runs an MCP server over the standard stdio transport, built on the official @modelcontextprotocol/sdk. It’s additive to the deterministic hook model neuron init wires for Claude Code and Codex CLI — for editors with no per-turn hook point (Cursor, Windsurf, Zed, Claude Desktop, Roo Code) — and available unconditionally to every client, hook-covered or not.

It takes no flags.

Tools exposed

Each tool is a thin wrapper over the same store methods the CLI itself calls — memory add’s write-time supersession gate, memory query’s relevance-gated search, exec’s pre-command lookup — never a parallel logic path.

Tool	Wraps	Input
neuron_remember	memory add	content (required), category, importance (1-5, default 3 — not inferred), supersedes, companion_of. No tags — server-inferred only
neuron_recall	memory query	query (required), categories. Returns { results, rejected }
neuron_query_exec	exec’s pre-command lookup	command_text (required). Lookup only — never spawns the command

No auth or scoping

A local stdio server is a subprocess your editor spawns directly, inheriting exactly the OS-level access any CLI invocation already has — no sandboxing layer exists in the MCP protocol itself.

Client config

No separate binary, no new bin entry — neuron init writes this automatically for detected harnesses (see neuron init), or configure a client by hand:

{
  "mcpServers": {
    "neuron": {
      "command": "npx",
      "args": ["-y", "@kovartravis/neuron", "mcp"]
    }
  }
}

Limitations

neuron mcp is best-effort, not deterministic. A live A/B test found agent-invoked neuron_recall with no hook backing it under-complied in every session tested — called only 5 of 8 times, and never fully compliant even when it was called. Ship it as reach for otherwise-unreachable editors, not as a guarantee equivalent to the deterministic hooks Claude Code and Codex CLI get.

Source: docs/design/rule-recall-ab/findings.md.